Unlike time series example, `condition` truncates forecast

numpyro
1

In the time series example, condition is used to condition the data for points at which we have observations and forecast at points where no data is available. However, if I use a batched model to do “time series” forecasting (i.e. no actual temporal dependence but I have exogenous regressors which vary over time. See this post.), the batch length is truncated to match the length of the data provided to condition. I had expected/hoped that I would get forecasts for any batched terms after the condition data. More, concretely:

def model1():
  normal = ny.sample("normal", dist.Normal(loc=np.array([-1, 1]), scale=1))
  print("normal", normal.shape)
  forecast = ny.deterministic("forecast", normal[-1, np.newaxis])
  print("forecast", forecast.shape)
def model2():
  normal_dist = dist.Normal(loc=np.array([-1, 1]), scale=1)   
  with ny.handlers.condition(data={"normal": np.array([3])}):
    normal = ny.sample("normal", normal_dist)
  print("normal", normal.shape)
def model3():
  normal_dist = dist.Normal(loc=np.array([-1, 1]), scale=1)   
  with ny.handlers.condition(data={"normal": np.array([3])}):
    normal = ny.sample("normal", normal_dist)
  print("normal", normal.shape)
  forecast = ny.sample("forecast", dist.Normal(loc=normal_dist.loc[-1], scale=normal_dist.scale))
  print("forecast", forecast.shape)

model1 prints out normal (2,); forecast (1,) as expected. But adding condition in model2 truncates normal to normal (1,).

Is the best solution to manually “reconstruct” a forecast as in model3?

Full notebook example can be found here.

2

This is not supported. For forecasting, we recommend splitting out observed statement and forecast statement by some if/else logic. Like

def model4():
  normal_dist = dist.Normal(loc=np.array([-1, 1]), scale=1)   
  with ny.handlers.condition(data={"normal": np.array([3])}):
    normal = ny.sample("normal", dist.Normal(normal_dist.loc[:1], normal_dist.scale)
  print("normal", normal.shape)
  forecast = ny.sample("forecast", dist.Normal(loc=normal_dist.loc[-1], scale=normal_dist.scale))
  print("forecast", forecast.shape)

Model 3 will give incorrect inference results because the observed data will be broadcasted to normal_dist.shape() when evaluating the log likelihood.

2 Likes
3

Ah, I didn’t realize that. Thanks!